Detection results from Claude Code, OpenCode, and Codex

[caffeinum]

Agent Detection Test Results

Asked each coding agent to fetch https://ai-docs-analytics-api.theisease.workers.dev/detect and report back.

---

Claude Code ✅

{
  "category": "coding-agent",
  "agent": "claude-code",
  "headers": {
    "user_agent": "axios/1.8.4",
    "accept": "text/markdown, text/html, */*"
  }
}

Detection signal: axios user-agent + text/markdown accept header

---

OpenCode ✅

{
  "category": "coding-agent",
  "agent": "opencode",
  "headers": {
    "user_agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/143.0.0.0 Safari/537.36",
    "accept": "text/plain;q=1.0, text/markdown;q=0.9, text/html;q=0.8, */*;q=0.1"
  }
}

Detection signal: q= weights in accept header (spoofed user-agent doesn't help them)

"so yeah, docs sites can fingerprint us via the accept header pattern even when user-agent is masked 🐱"

---

Codex 🤔

Using curl:

{
  "category": "bot",
  "agent": "curl",
  "filtered": true,
  "headers": {
    "user_agent": "curl/8.7.1",
    "accept": "*/*"
  }
}

Using search/browse tool:

{
  "category": "browsing-agent",
  "agent": "chatgpt-user",
  "filtered": true,
  "headers": {
    "user_agent": "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ChatGPT-User/1.0; +https://openai.com/bot",
    "accept": "text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.9"
  }
}

Note: Codex uses ChatGPT-User user-agent when browsing. Currently classified as browsing-agent (filtered). Need to determine if Codex CLI has a different signature vs ChatGPT browse mode.

---

Summary

Agent	Detected As	Signal	Status
Claude Code	claude-code	axios + text/markdown	✅ Working
OpenCode	opencode	text/markdown + q= weights	✅ Working
Codex (browse)	chatgpt-user	ChatGPT-User UA	🤔 Filtered as browsing-agent

TODO

• Investigate if Codex CLI (not browse) has a distinct signature
• Decide if ChatGPT-User should be coding-agent instead of browsing-agent

[caffeinum]

No rush on reclassifying ChatGPT-User - the two-table architecture means we can retroactively fix this anytime:

• ai_docs_raw_events stores original headers (immutable)
• ai_docs_visits stores classifications (can be reprocessed)

When we update detection logic, we can reprocess all raw events to regenerate visits with correct classifications.